$ the-wire · showcase
MILES resumes stop re-running startup eval, DeepEP lands in verl's uv images
By RepoJournal · Filed · About Alignment & Post-training · Composed from the cited sources · methodology
4 people shipped this
Two changes affect people who actually run these stacks: MILES no longer blocks a resumed run behind a fresh startup evaluation, and verl's uv Megatron extra now installs the DeepEP dispatcher it was documented to use.
[megatron, docker] fix: install DeepEP in uv training environments and restore NVSHMEM verl-project/verl
Megatron MoE training could enable the DeepEP dispatcher while the uv `megatron` extra left `deep_ep` uninstalled, and the CUDA 13 image lacked a system NVSHMEM at the path baked into the prebuilt wheel. The extra now pulls `deep-ep==1.2.1` from the matching Torch 2.13 / CUDA 13 wheelhouse for Linux x86_64 and aarch64, and the image installs NVSHMEM 3.4.5 with loader configuration.
Skip MILES startup evaluation on nonzero-step resumes allenai/open-instruct
With `skip_eval_before_train=false`, MILES Core used to rerun the full startup evaluation after checkpoint recovery; restoring update 60 blocked further training on another initial evaluation even though training had already begun. Startup evaluation now requires `start_rollout_id == 0`, so later resumes go straight to the normal periodic evaluation schedule.
[tool] feat: rl-insight trace weight-sync stages and report trainer metrics by step (#8315) verl-project/verl
Trainer scalars now go through rl-insight's `metric_log` when the installed package provides it, because `metric_gauge` carries no training step and its curves could not be plotted against `global_step`. Note the dependency: released 0.3.0 does not have `metric_log`, the step-aware API added in rl-insight#206.
Add named MILES inference presets and read-only configuration advice allenai/open-instruct
MILES learning examples used to cap inference at 64 requests and left operators copying hardware-specific serving knobs; they now ship named whole presets for the measured BF16 Olmo 3.5 geometry on dedicated TP1 B300/H100 engines, plus read-only planning advice. The node starter bounds unfinished generation to 1,536 samples independently of its six engines' reserved capacity, and a six-engine, ...
Add Claude as an advisory PR reviewer allenai/open-instruct
The remaining work is housekeeping: Ascend docs corrections, CI moves to writable config-lock and model-cache paths, the a3-8 and a3-16 runners swapped for 800i equivalents, and a new advisory Claude PR reviewer that always publishes as `COMMENT` so merge decisions stay with humans.
